The questions I want to work on.
My interests sit along one line: how molecular detail scales up into the behaviour of biological systems, and what computation can recover from that. Five connected themes, and thirteen proposals written out properly so there is something specific to argue with.
Where my interests sit.
Each theme carries the question I would actually want to answer within it, rather than a list of keywords.
Method choices, and why.
Start with what would falsify it
Before choosing a method I try to write down the result that would tell me I am wrong. If no such result exists, the question is not yet specified well enough to answer.
"What would this look like if the effect were not real?"
Prefer the method whose assumptions I can check
A simpler model whose assumptions are inspectable usually beats a more flexible one whose failure modes are invisible. Where I do use a flexible model, I want an applicability domain around it.
"Can I tell when this model is being asked something it cannot answer?"
Validate on something the model has never seen
Cross-validation measures how well a procedure generalises within one dataset. Independent data is what tests whether the finding survives contact with a different cohort, platform or batch.
"Does this replicate somewhere I did not tune it?"
Thirteen proposals.
Concepts I have written independently. None has been undertaken, none is supervised, and none has results. Each names the part I expect to be genuinely difficult, because that is usually the honest measure of whether a proposal has been thought through. Several grew out of problems I ran into in my own published work — annotation drift, pharmacogene coverage, CYP2D6 structural variation.